Generalized Ricci Bounds and Convergence of Metric Measure Spaces Bornes Généralisées de la Courbure Ricci et Convergence des Espaces Métriques Mesurés Karl-Theodor Sturm a a Institut für Angewandte Mathematik, Universität Bonn, Wegelerstrasse 6, 53125 Bonn, GERMANY Abstract We introduce and analyze curvature bounds Curv(M, d, m) K for metric measure spaces (M, d, m), based on convexity properties of the relative entropy Ent(. m). For Riemannian manifolds, Curv(M, d, m) K if and only if Ric M (ξ, ξ) K ξ 2 for all ξ T M. We define a complete separable metric D on the family of all isomorphism classes of normalized metric measure spaces. It has a natural interpretation in terms of mass transportation. Our lower curvature bounds are stable under D-convergence. We also prove that the family of normalized metric measure spaces with doubling constant C is closed under D-convergence. Moreover, the subfamily of spaces with diameter R is compact. To cite this article: K. T. Sturm, C. R. Acad. Sci. Paris, Ser. I 336 (2003). Résumé Nous introduisons et étudions bornes de la courbure Curv(M, d, m) K pour des espaces métriques mesurés (M, d, m), basées sur les propriétés de convexité de l entropie relative Ent(. m). Pour les varietés riemanniennes, Curv(M, d, m) K si et seulmement si Ric M (ξ, ξ) K ξ 2 pour tous ξ T M. Nous définissons une métrique D complete separable sur la famille des classes d isomorphie des espaces métrique mesurés normalisés. Elle a une naturel interprétation dans le contexte du transport de masse. Nos bornes inférieures de la courbure sont stables pour la D-convergence. Nous démontrons aussi que pour la D-convergence la famille des espaces métriques mesurés normalisés avec la constante de doublement C est fermée et, en plus, la sous-famille avec les diamètres R est compacte. Pour citer cet article : K. T. Sturm, C. R. Acad. Sci. Paris, Ser. I 336 (2003). A metric measure space is a triple (M, d, m) where (M, d) is a complete separable metric space and m is a measure on M (equipped with its Borel σ-algebra B(M)) which is locally finite in the sense that Email address: sturm@uni-bonn.de (Karl-Theodor Sturm). Preprint submitted to Elsevier Science 17th December 2004
m(b r (x)) < for all x M and all sufficiently small r > 0. A metric measure space (M, d, m) is called normalized iff m(m) = 1. Two metric measure spaces (M, d, m) and (M, d, m ) are called isomorphic iff there exists an isometry ψ : M 0 M 0 between the supports M 0 := supp[m] M and M 0 := supp[m ] M such that m = ψ m. The diameter of a metric measure space is given by diam(m, d, m) := sup{d(x, y) : x, y supp[m]}. Its variance is defined as Var(M, d, m) := inf d 2 M (z, x)dm (x) where the inf is taken over all metric measure spaces (M, d, m ) which are isomorphic to (M, d, m) and over all z M. The family of all isomorphism classes of normalized metric measure spaces with finite variances will be denoted by X 1, cf. [3], chapter 3 1 2. Definition 1 The distance D between two metric measure spaces is defined by ( D((M, d, m), (M, d, m )) = inf ˆd2 (x, y)d ˆm(x, y) ˆd, ˆm M M where the infimum is taken over all couplings ˆd of d, d and over all couplings ˆm of m, m Here a measure ˆm on the product space M M is called coupling of m and m iff ˆm(A M ) = m(a) and ˆm(M A ) = m (A ) for all measurable sets A M, A M ; a pseudo metric ˆd on the disjoint union M M is called coupling of d and d iff ˆd(x, y) = d(x, y) and ˆd(x, y ) = d (x, y ) for all x, y supp[m] M and all x, y supp[m ] M. One easily verifies that D((M, d, m), (M, d, m )) = inf ˆd W (ψ m, ψ m ) where the infimum is taken over all metric spaces ( ˆM, ˆd) with isometric embeddings ψ : M 0 ˆM, ψ : M 0 ˆM of the supports M 0 and M 0 of m and m, resp., and where ˆd W denotes the L 2 - Wasserstein distance between probability measures on ( ˆM, ˆd). In particular, if (M, d) = (M, d ) then D((M, d, m), (M, d, m )) d W (m, m ). In general, however, there will be no equality. Theorem 2 (X 1, D) is a complete separable length metric space. We say that a metric measure space (M, d, m) has the restricted doubling property with doubling constant C iff m(b 2r (x)) C m(b r (x)) for all x supp[m] and all r > 0. Theorem 3 The restricted doubling property is stable under D-convergence. That is, if for all n N the normalized metric measure spaces (M n, d n, m n ) have the restricted doubling property with a common doubling constant C and if (M n, d n, m n ) D (M, d, m) as n then also (M, d, m) has the restricted doubling property with the same constant C. Theorem 4 ( Compactness ) For each pair (C, R) R + R + the family X 1 (C, R) of all isomorphism classes of normalized metric measure spaces with ( restricted ) doubling constant C and diameter R is compact under D-convergence. Given a metric measure space (M, d, m) we denote by P 2 (M, d) the space of all probability measures ν on M with M d2 (o, x)dν(x) < for some (hence all) o M. The L 2 -Wasserstein distance on P 2 (M, d) is defined by d W (µ, ν) = inf( M M d2 (x, y)dq(x, y)) 1/2 with the infimum taken over all couplings q of µ and ν, see [7]. For ν P 2 (M, d) we define the relative entropy w.r.t. m by Ent(ν m) := lim ρ log ρ dm ɛ 0 {ρ>ɛ} if ν is absolutely continuous w.r.t. m with density ρ = dν dm and by Ent(ν m) := + if ν is singular w.r.t. m. Finally, we put P2 (M, d, m) := {ν P 2 (M, d) : Ent(ν m) < + }. 2 ) 1/2
Lemma 5 If m has finite mass, then Ent( m) is lower semicontinuous and on P 2 (M, d). Definition 6 (i) Given any number K R, we say that a metric measure space (M, d, m) has curvature K iff for each pair ν 0, ν 1 P 2 (M, d, m) there exists a geodesic Γ : [0, 1] P 2 (M, d, m) connecting ν 0 and ν 1 with Ent(Γ(t) m) (1 t)ent(γ(0) m) + tent(γ(1) m) K 2 t(1 t) d2 W (Γ(0), Γ(1)) for all t [0, 1]. Moreover, we put Curv(M, d, m) := sup{k R : (M, d, m) has curvature K.} (ii) We say that a metric measure space (M, d, m) has curvature K in the lax sense iff for each ɛ > 0 and for each pair ν 0, ν 1 P 2 (M, d, m) there exists an η P 2 (M, d, m) with d W (η, ν i ) 1 2 d W (ν 0, ν 1 ) + ɛ for each i = 0, 1 and Ent(η m) 1 2 Ent(ν 0 m) + 1 2 Ent(ν 1 m) K 8 d2 W (ν 0, ν 1 ) + ɛ. We denote the maximal K with this property by Curv lax (M, d, m). (iii) We say that a metric measure space (M, d, m) has locally curvature K if each point of M has a neighborhood M such that (M, d, m) has curvature K. The maximal K with this property will be denoted by Curv loc (M, d, m). Let us consider these curvature bounds under some of the Basic Transformations: Isomorphisms. Curv(M, d, m) = Curv(M, d, m ) for each (M, d, m ) isomorphic to (M, d, m); Scaling. Curv(M, α d, βm) = α 2 Curv(M, d, m) for all α, β > 0; Weights. Curv(M, d, e V m) Curv(M, d, m) + HessV for each lower bounded, measurable function V : M R where HessV := sup{k R : V is K-convex on supp[m]}; Subsets. Curv(M, d, m) Curv(M, d, m) for each convex M M; Products. Curv(M, d, m) = inf i {1,...,l} Curv(M i, d i, m i ) if (M, d, m) = l i=1 (M i, d i, m i ) and if M is nonbranching and compact. Here a metric space (M, d) is called nonbranching iff for each quadruple of points z, x 0, x 1, x 2 with z being the midpoint of x 0 and x 1 as well as the midpoint of x 0 and x 2 it follows that x 1 = x 2. Theorem 7 Let M be a complete Riemannian manifold with Riemannian distance d and Riemannian volume m and put m = e V m with a C 2 function V : M R. Then Curv(M, d, m ) = inf {Ric M (ξ, ξ) + HessV (ξ, ξ) : ξ T M, ξ = 1}. In particular, (M, d, m) has curvature K if and only if the Ricci curvature of M is K. See [4] for the case V = 0 or [5] for the general case. Note that in the above Riemannian setting for each pair of points ν 0, ν 1 in P 2 (M, d, m) there exists a unique geodesic connecting them [1]. Lemma 8 If M is compact then Curv(M, d, m) = Curv lax (M, d, m). Of fundamental importance is that our curvature bounds for metric measure spaces are stable under convergence and that local curvature bounds imply global curvature bounds. The latter is in the spirit of the Globalization Theorem of Topogonov for lower curvature bounds (in the sense of Alexandrov) for metric spaces. Theorem 9 Let (M, d, m) be a compact, nonbranching metric measure space such that P 2 (M, d, m) is a geodesic space. Then Curv(M, d, m) = Curv loc (M, d, m). 3
Theorem 10 Let ((M n, d n, m n )) n N be a sequence of normalized metric measure spaces with uniformly D bounded diameter and with (M n, d n, m n ) (M, d, m). Then lim sup n Curv lax (M n, d n, m n ) Curv lax (M, d, m). Example 1 Given any abstract Wiener space (M, H, m) define a pseudo metric on M by d(x, y) := x y H if x y H and d(x, y) := else and consider the pseudo metric measure space (M, d, m). Then Curv lax (M, d, m) = 1. Of course, formally this does not fit in our framework. Nevertheless, the definition of the L 2 -Wasserstein distance d W derived from this pseudo metric d perfectly makes sense (cf. [2]) and also the relative entropy is well-defined. Lower bounds for the curvature will imply upper estimates for the volume growth of concentric balls. In the Riemannian setting, this is the content of the famous Bishop-Gromov volume comparison theorem. In the general case (without any dimensional restriction) these estimates, however, have to take into account that the volume can grow much faster than exponentially. For instance, we already observe squared exponential volume growth if we equip the one-dimensional Euclidean space with the measure dm(x) = exp( Kx 2 /2) dx for some K < 0. Theorem 11 Let (M, d, m) be an arbitrary metric measure space with Curv(M, d, m) K for some K 0. For fixed x supp[m] M consider the volume growth v R := m(b R (x)) of closed balls centered at x. Then for all R 2ɛ > 0 ( ) v R v 2ɛ (v 2ɛ /v ɛ ) R/ɛ exp K (R + ɛ/2) 2 /2. In particular, each ball in M has finite volume. Theorem 12 If Curv(M, d, m) K 0 then for all x M and for all R 3ɛ > 0 the volume of spherical shells v R,ɛ := m(b R (x) \ B R ɛ (x)) can be estimated by v R,ɛ v 3ɛ (v 3ɛ /v ɛ ) R/2ɛ exp ( K [ (R 3ɛ) 2 ɛ 2] /2 ). In particular, K > 0 implies that m has finite mass and finite variance. undefined References [1] D. Cordero-Erausquin, R. McCann, M. Schmuckenschläger (2001): A Riemannian interpolation inequality à la Borell, Brascamb and Lieb. Invent. Math., 146, 219-257. [2] D. Feyel, A. S. Üstünel (2004): Monge-Kantorovitch measure transportation and Monge-Ampère equation on Wiener space. Probab. Theory Related Fields 128, no. 3, 347 385. [3] M. Gromov (1999): Metric structures for Riemannian and non-riemannian spaces. With appendices by M. Katz, P. Pansu and S. Semmes. Birkhäuser Boston Inc., Boston, MA. 4
[4] M.-K. von Renesse, K. T. Sturm (2003): Transport inequalities, gradient estimates, entropy and Ricci curvature. To appear in Comm. Pure Appl. Math. [5] K. T. Sturm (2004): Convex functionals of probability measures and nonlinear diffusions on manifolds. Preprint 2004. [6] K. T. Sturm (2005): On the geometry of metric measure spaces. In preparation. [7] C. Villani (2003) Topics in Mass Transportation. Graduate Studies in Mathematics. American Mathematical Society. 5